Progress and Trends of Deep Learning in CT Image Lesion Area Detection

Deep learning has achieved some good results in the field of medical image analysis to date, but most existing models for CT lesion detection have been developed for a single organ and lack a cross-site universal framework; in addition, their detection performance is highly dependent on specific clinical conditions. The following is an overview of the current stage, development and application cases of deep learning detection methods for four types of CT images: pulmonary nodules, brain lesions, liver lesions and bone lesions. The problems in Model Design and how to solve them are presented in this paper. Based on the comparison of the quantitative results and technical attributes of the main models, it can be observed that the general direction of development is moving towards three-dimensional representations, multiple tasks, integration of clinical knowledge, and so on. At the same time, the main bottlenecks in current research have also been identified in the three areas of data, clinical diagnosis and algorithms; high annotation costs, poor interpretability, insufficient use of three-dimensional information and weak generalisation ability are among them. This paper points out that there are deficiencies in assessing the strengths and weaknesses of methods that are not tied to particular situations. According to the kind of anatomical place and clinical conditions, a specific detection plan will be chosen. In the future, semi-supervised learning, multi-task integrated models and interpretable AI will be employed to increase the accuracy and clinical applicability of this field.